On the Navier-Stokes equations and the Hamilton-Jacobi-Bellman equation on the group of volume preserving diffeomorphisms

Abstract

In this paper, we give a new derivation of the incompressible Navier-Stokes equations on a compact Riemannian manifold $M$ via the Bellman dynamic programming principle on the infinite dimensional group of volume preserving diffeomorphisms $G={\rm SDiff}(M)$. In particular, when the viscosity vanishes, we give a new derivation of the incompressible Euler equation on a compact Riemannian manifold. The main result of this paper indicates an interesting relationship between the incompressible Navier-Stokes equations on $M$ and the Hamilton-Jacobi-Bellman equation on $G={\rm SDiff}(M)$.

MSC2010 Classification: primary 35Q30, 49L20，secondary 58J65, 60H30

Keywords: Navier-Stokes equations, Bellman dynamic programming principle, Hamilton-Jacobi-Bellman equation, the group of volume preserving diffeomorphisms

1 Introduction

Hydrodynamics is one of the fundamental areas in mathematics and physics. In [4] , Arnold gave a geometric interpretation of the incompressible Euler equation on a Riemannian manifold $M$ as a geodesic on the group $G={\rm SDiff}(M)$ of volume preserving diffeomorphisms equipped with the $L^{2}$-right invariant Riemannian metric. See also [5]. In [14] (see also Taylor [35]), the local existence and uniqueness of the Cauchy problem to the incompressible Euler equation (i.e., Arnold’s geodesic) within suitable Sobolev space was well-established. When ${\rm{dim}\hskip 0.85358pt}M=2$, the global existence and uniqueness is known. See e.g. [35]. When ${\rm{dim}\hskip 0.85358pt}M\geq 3$, it remains a major outstanding problem whether the Euler equation has global smooth solutions. On the other hand, concerning the minimal geodesic linking two points in $G$, Ebin and Marsden [14] proved that if $h$ belongs to a sufficiently small neighborhood of the identity map for a suitable Sobolev norm, there is a unique minimal geodesic connecting $h$ and the identity map. However, a striking result of Shnirelman [34] shows that, for $M=[0,1]^{3}$ and for a large class of data, there is no minimal geodesic. This shows the difficulty of applying the variational method to construct nonstationary incompressible flows in the three-dimensional case. In recent years, there have been some new investigations on the incompressible Euler equation. For instance, Brenier [6] introduced the notion of generalized solutions to the imcompressible Euler equation and proved the existence of the generalized solutions. He also proved a regularity estimate on the pressure of the generalized solution to the Euler equation. In [1], Ambrosio and Figalli improved Brenier’s regularity estimates for the gradient of the pressure.

When the fluid is viscid, the incompressible Navier-Stokes equation on $\mathbb{R}^{d}$ reads

$$\begin{cases}$$\partial_{t}u+\left(u\cdot\nabla\right)u-\nu\Delta u=-\nabla p,$$&\mbox{$$}\\ $$\rm{div}\hskip 0.85358ptu=0,$$&\mbox{$$}\end{cases}$$ (1.1)

where $\nu>0$ is the viscosity constant. In this situation, the system is no longer conservative because of the friction term. Local existence with initial data $u_{0}\in L^{n}(\mathbb{R}^{d})$ was established in Kato [24]. Ebin and Marsden [14] proved the local existence and uniqueness of smooth solutions to the incompressible Navier-Stokes equation on Riemannian manifolds. See also Taylor [35]. When ${\rm{dim}\hskip 0.85358pt}M=2$, global existence of smooth solutions is also known. When ${\rm{dim}\hskip 0.85358pt}M\geq 3$, it is a long time outstanding open problem (one of the Millennium Prize Problems) whether the incompressible Navier-Stokes equation has global smooth solutions.

It is interesting to ask the question whether one can extend Arnold’s point of view to give a new derivation of the incompressible Navier-Stokes equation $(\ref{NS1})$. It is well-known that the Laplacian $\Delta$ is the infinitesimal generator of the Brownian motion. So, it is reasonable to derive the incompressible Navier-Stokes equations by adding the effect of the Brownian motion to the Euler equation. In some sense, this gives an explanation of the physical term “internal friction” (Landau-Lifshitz [26], see also Serrin [33]) by a probabilistic point of view. To our knowledge, the connections between Navier-Stokes equations and stochastic evolution traced back to Chorin [8], and a stochastic Hamiltonian approach was given by Inoue and Funaki [23]. Since then, many people have studied the incompressible Navier-Stokes equation using a stochastic Lagrangian variational calculus together with Nelson’s stochastic mechanics [32]. For this, we mention the works by Yasue [36, 37], Nakagomi-Yasue-Zambrini[31] and Esposito et al [16].

In recent years, more probabilistic works have been developed for the incompressible Navier-Stokes equations. In [28], Le Jan and Sznitman used a backward-in-time branching process in Fourier space to express the velocity field of a three-dimensional viscous fluid as the average of a stochastic process, which then leads to a new existence theorem. In [7], Busnello introduced a probabilistic approach to the existence of a unique global solution for two dimensional Navier-Stokes equations. In [8], Busnello, Flandoli and Remito gave the probabilistic representation formulas for the vorticity and the velocity of three dimensional Navier-Stokes equations and gave a new proof of the local existence of solution. In [10], Cipriano and Cruzeiro gave a stochastic variational principle for two dimensional incompressible Navier-Stokes equations by using the Brownian motions on the group of homeomorphisms on the torus. In [13], Cruzeiro and Shamarova established a connection between the strong solution to the spatially periodic Navier-Stokes equations and a solution to a system of forward-backward stochastic differential equations on the group of volume preserving diffeomorphisms of a flat torus. They also constructed representations of the strong solution to the Navier-Stokes equations in terms of diffusion processes. In [11], Constantin and Iyer derived a probabilistic representation of the three-dimensional Navier-Stokes equations based on stochastic Lagrangian paths. As they proved, the particle trajectories obey SDEs driven by a uniform Wiener process, the inviscid Weber formula for the Euler equations of ideal fluids is used to recover the velocity field. This method admits a self-contained proof of local existence for the nonlinear stochastic system and can be extended to formulate stochastic representations of related hydrodynamic-type equations, including viscous Burgers equations and Lagrangian-averaged Navier-Stokes alpha models. By reversing the time variable, Zhang [40] derived a stochastic representation for backward incompressible Navier-Stokes equations in terms of stochastic Lagrangian paths and gave a self-contained proof of local existence of solutions in Sobolev spaces in the whole space. He also gave an alternative proof to the global existence for the two-dimensional incompressible Navier-Stokes equations with large viscosity. For extension of the above results to compact Riemannian manifolds, see Fang-Zhang[20], Arnaudon-Cruzeiro[2], Arnaudon-Cruzeiro-Fang [3], Fang-Luo[19], Luo[29] and Fang [18]. Due to the limit of the paper, we will not try to give a full description and citation of all previous works on the probabilistic studies for incompressible Navier-Stokes equations.

The purpose of this paper is to use the Bellman dynamic programming principle on the infinite dimensional group of volume prerserving diffeomorphisms on a compact Riemannian manifold to derive the incompressible Navier-Stokes equation on a compact Riemannian manifold. To the best knowledge of the authors, this is a new point of view even in the case where the viscosity coefficient vanishes. That is to say, the incompressible Euler equation can be derived from the deterministic dynamic programming principle on the infinite dimensional group of volume preserving diffeomorphisms over a compact Riemannian manifold.

Our results suggest a deep and close connection between the incompressible Navier-Stokes equations with the Bellman dynamic programming principle on the group of volume preserving diffeomorphisms. We hope that our work may bring some new point of views for the further study of the incompressible Navier-Stokes and Euler equations.

The rest of the paper is organized as follows. In Section 2, we extend the Bellman dynamic programming principle and derive the Hamilton-Jacobi-Bellman equation on compact Riemannian manifolds. In Section 3, we prove the Bellman dynamic programming principle on the infinite dimensional group of volume preserving diffeomorphisms $G={\rm SDiff}(M)$. In Section 4, we derive the incompressible Navier-Stokes equations on a compact Riemannian manifold using the Bellman dynamic programming principle and the Hamilton-Jacobi-Bellman equation on $G={\rm SDiff}(M)$. In Section 5, we give a new derivation of the incompressible Euler equation via the deterministic dynamic programming principle and the Hamilton-Jacobi equation on $G={\rm SDiff}(M)$. In Section 6, we formulate our main result which indicates an interesting relationship between the incompressible Navier-Stokes equations on $M$ and the Hamilton-Jacobi-Bellman equation on the group of volume preserving diffeomorphisms $G={\rm SDiff}(M)$, and we raise some problems for further research work.

2 Dynamic programming principle and HJB equation on Riemannian manifolds

In this section, we extend the classical Bellman dynamic programming principle from the Euclidean space to compact Riemannian manifolds, and derive the Hamilton-Jacobi-Bellman and Burgers equations on Riemannian manifolds.

First, for the convenience of the reader, we briefly review the standard results on the Bellman dynamic programming principle and the Hamilton-Jacobi-Bellman equation in stochastic control theory.

Consider the following controlled SDE on $\mathbb{R}^{d}$

$$\begin{cases}$$dx^{u}(t)=\sqrt{2\nu}dB_{t}+u(t,x^{u}(t))dt,$$&\mbox{$$}\\ $$x^{u}(0)=x,$$&\mbox{$$}\end{cases}$$ (2.1)

where $B_{t}$ is a $\mathbb{R}^{d}$-valued standard Brownian motion. The corresponding Lagrangian action on $\mathbb{R}^{d}$ is given by

$$S[x^{u}]=\mathbb{E}\left[\int_{0}^{T}\frac{1}{2}\|D_{t}x^{u}(t)\|_{\mathbb{R}^{d}}dt\right].$$ (2.2)

Let $\Psi\in C(\mathbb{R}^{d},\mathbb{R})$, and define the value function

$$W(t,x)=\mathbb{E}\left[\left.\int_{t}^{T}\frac{1}{2}\|D_{s}x^{u}(s)\|_{\mathbb{R}^{d}}^{2}ds+\Psi(x^{u}(T))\right|x^{u}(t)=x\right].$$ (2.3)

By Bellman’s principle of stochastic dynamic programming on Euclidean space, it is well-known that $W$ satisfies the Hamilton-Jacobi-Bellman equation

$$\partial_{t}W+\nu\Delta W-\frac{1}{2}\|\nabla W\|_{\mathbb{R}^{d}}^{2}=0,$$ (2.4)

and the optimal Markov control is given by

$$u(t,x)=-\nabla W(t,x).$$ (2.5)

See e.g. Zambrini [39], Fleming and Soner [21]. Suppose that $W\in C^{1,3}([0,T]\times\mathbb{R}^{d},\mathbb{R})$. Differentiating the both sides of $(\ref{HJB-1})$ with respect to spatial variable $x\in\mathbb{R}^{d}$, we derive the Burgers equation on $\mathbb{R}^{d}$

$$\partial_{t}u+\nu\Delta u+u\cdot\nabla u=0.$$ (2.6)

2.1 Dynamic programming principle on Riemannian manifolds

Let $M$ be a $d$-dimensional compact Riemannian manifold without boundary, $(\Omega,\mathcal{F},(\mathcal{F}_{s}),\mathbb{P})$ be a complete probability space endowed with an increasing filtration $(\mathcal{F}_{s})$ which satisfies the usual condition.

Let $u\in C^{1,2}([0,T]\times M,TM)$. Let $x_{0}\in M$ be fixed, and $B_{t}$ be a standard Brownian motion on $T_{x_{0}}M$. By Itô’s SDE theory, the following controlled stochastic differential equation on $M$

$$\begin{cases}$$dx(s)=\sqrt{2\nu}U_{t\rightarrow s}\circ dB_{s}+u(s,x(s))ds,$$&\mbox{$t\leq s\leq T$}\\ $$x(t)=x,$$&\mbox{$x\in M$,}\end{cases}$$ (2.7)

admits a unique strong solution, where $U_{t\rightarrow s}:T_{x(t)}M\rightarrow T_{x(s)}M$ denotes the stochastic parallel transport along the trajectory of the diffusion process $\{x(r):r\in[t,s]\}$, which satisfies the covariant SDE on $M$

$$\nabla_{\circ dx(s)}U_{t\rightarrow s}=0,\ \ \ U_{t\rightarrow t}={\rm id}_{T_{x(t)}M},\ \ \forall\ t\leq s\leq T.$$ (2.8)

where $\circ$ denotes the Stratonovich differentiation. See [30, 15].

Let $\mathcal{U}_{ad}$ be the set of all time dependent adapted smooth vector fields $X$ on $M$ which satisfies the $L^{2}$-integrability condition

$$\|X\|_{L^{2}}^{2}=\mathbb{E}\left[\int_{0}^{T}|X(s,x(s))|_{T_{x(s)}M}^{2}ds\right]<+\infty.$$ (2.9)

Given a Lagrangian function $L\in C^{1}([0,T]\times TM,\mathbb{R})$ and $\Psi\in C(M,\mathbb{R})$, we consider the following minimization problem on a finite time interval $t\leq s\leq T$

$${\rm minimizes}\ J(t,x;u)=\mathbb{E}_{t,x}\left[\int_{t}^{T}L(s,x(s),u(s))ds+\Psi(x(T))\right],$$ (2.10)

subject to $u\in\mathcal{U}_{ad}=\left\{u{\rm{~{}is~{}(\mathcal{F}_{s})~{}adapted~{}}}:\mathbb{E}\left[\int_{0}^{T}|u(s,x(s))|_{T_{x(s)}M}^{2}ds\right]<\infty\right\}$. For simplicity, we write $\mathbb{E}_{t,x}[\cdots]$ instead of $\mathbb{E}\left[\left.\cdots\right|x(t)=x\right]$ throughout this paper.

By the same argument as in the proof of the dynamic programming equation on the Euclidean spaces, see e.g. Fleming and Soner[21], we can prove the following

Theorem 2.1

Let $M$ be a compact Riemannian manifold without boundary. Let $W$ be the value function defined by

$$W(t,x)=\inf\limits_{u\in\mathcal{U}_{ad}}J(t,x;u).$$ (2.11)

Assume that $L\in C^{1}([0,T]\times TM,\mathbb{R})$ and $W\in C^{1,2}([0,T]\times M,\mathbb{R})$. Then the dynamic programming equation holds

$$0=\min\limits_{v\in\mathcal{U}}\left[A^{v}W(t,x)+L(t,x,v)\right].$$ (2.12)

Proof. Taking the constant control $u(s)=v$ in a small time interval $s\in[t,t+h]$, we have

$$W(t,x)\leq\mathbb{E}_{t,x}\int_{t}^{t+h}L(s,x(s),v)ds+\mathbb{E}_{t,x}W(t+h,x(t+h)).$$ (2.13)

Subtracting first $W(t,x)$ from both sides of $(\ref{lopt})$, then dividing by positive constant $h$ and letting $h$ tend to $0$. Using the continuity of $L$, we have

$$\lim\limits_{h\rightarrow 0^{+}}h^{-1}\mathbb{E}_{t,x}\int_{t}^{t+h}L(s,x(s),v)ds=L(t,x,v).$$

Using the covariant Itô-Dynkin formula (see [15]) and similar to the proof on the Euclidean space as in [21], if $W\in C^{1,2}([0,T]\times M,\mathbb{R})$, we have

$$\lim\limits_{h\rightarrow 0^{+}}h^{-1}\left[\mathbb{E}_{tx}W(t+h,x(t+h))-W(t,x)\right]$$

$$= \lim\limits_{h\rightarrow 0^{+}}h^{-1}\mathbb{E}_{tx}\int_{t}^{t+h}A^{v}W(s,x(s))ds=A^{v}W(t,x),$$

where $A^{v}=\partial_{t}+\nu\Delta+v\cdot\nabla$. Therefore, for all constant control $v\in\mathcal{U}_{ad}$, we obtain

$$0\leq A^{v}W(t,x)+L(t,x,v).$$ (2.14)

On the other hand, if $u^{\ast}$ is the optimal Markov control, we have

$$W(t,x)=\mathbb{E}_{t,x}\int_{t}^{t+h}L(s,x^{\ast}(s),u^{\ast}(s,x^{\ast}(s)))ds+\mathbb{E}_{t,x}W(t+h,x^{\ast}(t+h)),$$

where $x^{\ast}(s)$ is the diffusion process generated by $A^{u^{\ast}}$. If $L\in C^{1}([0,T]\times TM,\mathbb{R})$ and $W\in C^{1,2}([0,T]\times M,\mathbb{R})$, a similar argument as for the proof on the Euclidean space (see e.g. [21]) gives

$$0=A^{u^{\ast}}W(t,x)+L(t,x,u^{\ast}(t,x)).$$ (2.15)

Combining (2.14) and (2.15), we derive the dynamic programming equation on $M$

$$0=\min\limits_{v\in\mathcal{U}}\left[A^{v}W(t,x)+L(t,x,v)\right].$$

The value function $W$ satisfies the terminal condition

$$W(T,x)=\Psi(x).$$

The above argument suggests that an optimal Markov control satisfies

$$u^{\ast}(t,x)\in\arg\min\left[A^{v}W(t,x)+L(t,x,v)\right],$$

where $\arg\min f:=\left\{u\in\mathcal{U}_{ad}:~{}f(u)\leq f(v),~{}\forall v\in\mathcal{U}_{ad}\right\}$.   

2.2 The Hamilton-Jacobi-Bellman equation on Riemannian manifolds

In this subsection, we derive the Hamilton-Jacobi-Bellman equation from the Bellman dynamic programming principle on compact Riemannian manifolds.

Let $(M,g)$ be a compact Riemannian manifold, $g$ the Riemannian metric, and $d\mu=\sqrt{{\rm det}g(x)}dx$ the volume measure on $(M,g)$. Let $\nabla^{TM}$ be the Levi-Civita covariant derivative operator on $TM$, and $\nabla$ (respectively, $\Delta$) the Riemannian gradient (respectively, the Laplace-Beltrami) operator on $(M,g)$.

Following [32], see also [10, 2, 3], the Nelson derivative of the diffusion process $x(t)$ on $M$ is defined as follows

$$D_{t}x(t):=\lim\limits_{\varepsilon\rightarrow 0}\mathbb{E}\left[\left.{U_{t+\varepsilon\rightarrow t}dx(t+\varepsilon)-dx(t)\over\varepsilon}\right|\mathcal{F}_{t}\right].$$ (2.16)

Note that, by [32, 3], for $x(t)$ defined by SDE $(\ref{SDEM})$, the Nelson derivative of $x(t)$ is given by

$$D_{t}x(t)=u(t,x(t)).$$

We now state the main result of this subsection.

Theorem 2.2

Given $V\in C^{1}(M,\mathbb{R})$ and $\Psi\in C(M,\mathbb{R})$. Let

$$L(s,x(s),u(s))=\frac{1}{2}|u(s,x(s))|_{T_{x(s)}M}^{2}-V(x(s))$$ (2.17)

be the Lagrangian function, and let

$$W(t,x)=\mathbb{E}_{t,x}\left[\int_{t}^{T}L(s,x(s),u(s))ds+\Psi(x(T))\right]$$

be the value function. Suppose that $W\in C^{1,3}([0,T]\times M,\mathbb{R})$. Then $W$ satisfies the Hamilton-Jacobi-Bellman equation on $M$

$$\partial_{t}W+\nu\Delta W-\frac{1}{2}|\nabla W|^{2}_{T_{x}M}-V(x)=0,$$ (2.18)

with the terminal condition $W(T,x)=\Psi(x)$. Moreover, the optimal control is given by

$$u^{*}(t,x)=-\nabla W(t,x).$$ (2.19)

Define

$$v(t,x)=\nabla W(T-t,x).$$ (2.20)

Then $v$ satisfies the Burgers equation on $M$

$$\partial_{t}v+\nu\square v+\nabla_{v}^{TM}v+\nabla V=0,$$ (2.21)

where $\square=-\Delta+{\rm Ric}$ is the Hodge Laplacian, $\Delta={\rm Tr}\nabla^{2}$ is the covariant Laplacian, ${\rm Ric}$ is the Ricci curvature on $M$, and $\nabla^{TM}$ denotes the Levi-Civita covariant derivative operator on $TM$.

Proof. The Bellman stochastic dynamic programming principle on $M$ says that the value function $W$ satisfies the corresponding dynamic programming equation

$$0=\inf_{v\in T_{x}M}\left\{\partial_{t}W+\nu\Delta W+\langle v,\nabla W\rangle+\frac{1}{2}|v|_{T_{x}M}^{2}-V(x)\right\}.$$ (2.22)

Using the Riemannian inner product on the tangent space, we have

$$\inf_{v\in\mathcal{U}}\left\{\partial_{t}W+\nu\Delta W+v\cdot\nabla W+\frac{1}{2}|v|^{2}_{T_{x}M}-V(x)\right\}$$

$$= \partial_{t}W+\nu\Delta W+\inf_{v\in\mathcal{U}}\left\{v\cdot\nabla W+\frac{1}{2}|v|^{2}_{T_{x}M}\right\}-V(x)$$

$$= \partial_{t}W+\nu\Delta W-\frac{1}{2}|\nabla W|^{2}_{T_{x}M}-V(x).$$

This proves that $W$ satisfies the Hamilton-Jacobi-Bellman equation $(\ref{ch2-eq:9})$ on $M$. Moreover, from the above argument, we see that the optimal Markov control $u^{*}$ is given by

$$u^{*}=-\nabla W.$$

Assuming that $W\in C^{1,3}([0,T]\times M)$, and taking the covariant differentiation on both sides of the Hamilton-Jacobi-Bellman equation $(\ref{ch2-eq:9})$, we have

$$\partial_{t}\nabla W+\nu\nabla\Delta W-\frac{1}{2}\nabla|\nabla W|^{2}-\nabla V(x)=0.$$ (2.23)

Using the commutation formula in differential geometry

$$\Delta\nabla W=\nabla\Delta W+{\rm Ric}(\nabla W),$$ (2.24)

and the geometric formula

$${1\over 2}\nabla|\nabla W|^{2}=\langle\nabla\nabla W,\nabla W\rangle=\nabla^{TM}_{v}v,$$ (2.25)

we derive that $v(t,x)=\nabla W(T-t,x)$ satisfies the Burgers equation $(\ref{M-Burgers})$.   

2.3 Regularity of solution to HJB equation on $M$

Note that $(\ref{M-Burgers})$ is a system of quasi-linear parabolic partial differential equations on compact Riemannian manifolds. By standard argument as used in [35] on quasi-linear parabolic PDEs, we can prove the local existence and uniqueness of the Cauchy problem to the Burgers equation within suitable Sobolev space on a compact Riemannian manifold $M$.

Under the Cole-Hopf transformation

$$\Phi(t,x)=\exp{\left[-{W(t,x)\over{2\nu}}\right]},~{}~{}\varphi(x)=\exp[-\Psi(x)],$$ (2.26)

the Hamilton-Jacobi-Bellman equation $(\ref{ch2-eq:9})$ becomes the heat equation

$$\partial_{t}\Phi+\nu\Delta\Phi+\frac{1}{2\nu}V(x)\Phi=0,\ \ \ \Phi(T,x)=\varphi(x),$$ (2.27)

which is the backward heat equation with a potential term $\frac{1}{2\nu}V\Phi$.

Conversely, we can use the inverse of the Cole-Hopf transformation $(\ref{Hopf})$ to derive the Hamilton-Jacobi-Bellman equation $(\ref{ch2-eq:9})$ from the heat equation $(\ref{HeatV})$ provided that $\Phi$ is regular. Using the Feynman-Kac formula for Eq. $(\ref{HeatV})$, it is well-known that $\Phi\in C^{1,2}([0,T]\times M)$ provided that $\varphi\in C^{2}(M)$. Indeed, we have the following

Theorem 2.3

Suppose that the terminal value function $x\mapsto\Psi(x)$ is $C^{k}$-smooth. Then the value function $x\mapsto W(t,x)$ is also $C^{k}$-smooth on $M$.

Proof. The proof is standard and is omitted.   

3 Dynamic programming principle on $G={\rm SDiff}(M)$

In this section, we prove the Bellman dynamic programming principle on the group of volume preserving diffeomorphisms.

Let $(\Omega,\mathcal{F},(\mathcal{F}_{s}),\mathbb{P})$ be a filtered probability space satisfying the usual condition. Let $M$ be a $d$-dimensional compact Riemannian manifold, $G={\rm SDiff}(M)$ the group of volume preserving diffeomorphisms on $M$, and $\mathcal{G}$ its Lie algebra, i.e.,

$$\mathcal{G}=\left\{X\in\mathcal{X}(M):{\rm div}\hskip 0.56905ptX=0\right\}$$

is set of divergence free smooth vector fields on $M$.

Let $Q$ be a symmetric non-negative Hilbert-Schmidt operator on $\mathcal{G}$ equipped with the $L^{2}$-Riemannian metric defined in Section $2$. Recall the following

Definition 3.1

A continuous $\mathcal{G}$-valued stochastic process $B^{\mathcal{G}}_{s}$ is called a $Q$-Wiener process or $Q$-Brownian motion if

(1) $B^{\mathcal{G}}_{s}=id$ a.s.,

(2) $B^{\mathcal{G}}_{s}$ has independent increments $B^{\mathcal{G}}_{st}=B^{\mathcal{G}}_{t}-B^{\mathcal{G}}_{s}$ for all $0\leq s<t$,

(3) the increments have the following Gaussian laws: $\mathbb{P}\circ(B^{\mathcal{G}}_{st})^{-1}=\mathcal{N}(0,(t-s)Q)$ for all $0\leqslant s\leqslant t$.

Let $\{g(s),s\in[t,T]\}$ be defined as the solution to the following Stratonovich stochastic differential equation on $G$

$$\begin{cases}$$dg(s)=b(g(s),u(s))ds+\sigma(g(s),u(s))\circ dB^{\mathcal{G}}_{s},$$&\mbox{}\\ $$g(t)=g,$$&\mbox{}\end{cases}$$ (3.1)

where $b:TG\rightarrow TG$, $\sigma:TG\rightarrow\mathcal{L}(TG;TG)$ satisfy the Lipschitz condition, $TG$ is the tangent bundle on $G$, $\mathcal{L}(TG;TG)$ is the bounded linear operator, and $u(s)$ is a $T_{g(s)}G$-valued $\mathcal{F}_{s}$-adapted process along the diffusion process $g(s)$ on $G$. We call such $u$ an admissible control on SDE $(\ref{SDEG-1})$ on $G$.

Following [32], see also [10, 2, 3], the Nelson derivative of the diffusion process $g(t)$ on $G$ is defined as follows

$$D_{t}g(t):=\lim\limits_{\varepsilon\rightarrow 0}\mathbb{E}\left[\left.{//^{g}_{t+\varepsilon\rightarrow t}dg(t+\varepsilon)-dg(t)\over\varepsilon}\right|\mathcal{F}_{t}\right]$$

where $//^{g}_{t+\varepsilon\rightarrow t}:T_{g(t+\varepsilon)}G\rightarrow T_{g(t)}G$ denotes the stochastic parallel transport from $T_{g(t+\varepsilon)}G$ to $T_{g(t)}G$. Indeed, it is well-known that

$$D_{t}g(t)=b(g(t),u(t)).$$

Let

$$\mathcal{U}=\{u:[0,T]\times G\rightarrow TG\ \ {\rm adapted\ to}\ (\mathcal{F}_{s}):u(s,g(s))\in T_{g(s)}G\}$$

be the set of admissible controls on $G$. For simplicity, we use $\mathbb{E}_{t,g}[\cdots]$ instead of $\mathbb{E}\left[\left.\cdots\right|g(t)=g\right].$

We introduce the value function on $[t,T]\times G$ as follows

$$W(t,g)=\inf\limits_{u\in\mathcal{U}}\mathbb{E}_{t,g}\left[\int_{t}^{T}L(s,g(s),u(s))ds+\Psi(g(T))\right].$$ (3.2)

The following result extends the Bellman dynamic programming principle from Euclidean spaces to the infinite dimensional group of volume preserving diffeomorphisms.

Theorem 3.2

(Bellman’s principle of dynamic programming on $G$) For $h$ small enough such that $0\leq h\leq T-t$, we have

$$W(t,g)=\inf\limits_{u\in\mathcal{U}}\mathbb{E}_{t,g}\left[\int_{t}^{t+h}L(s,g(s),u(s))ds+W(t+h,g(t+h))\right].$$ (3.3)

Proof. We first prove the “$\leq$” part. Choose any control $\tilde{u}_{1}(\cdot)\in\mathcal{U}$, and let $g_{1}(\cdot)$ be the strong solution of the stochastic differential equation

$$\begin{cases}$$dg_{1}(s)=b(g_{1}(s),u_{1}(s))ds+\sigma(g_{1}(s),u_{1}(s))dB^{G}(s),$$&\mbox{}\\ $$g_{1}(t)=g.$$&\mbox{}\end{cases}$$ (3.4)

By definition of the infimum, there exists $u_{2}(\cdot)\in\mathcal{U}$ such that

$$W(t+h,g_{1}(t+h))+\varepsilon\geq\mathbb{E}_{t,g}\left[\int_{t+h}^{T}L(s,g_{2}(s),u_{2}(s))ds+\Psi(g_{2}(T))\right],$$

where

$$\begin{cases}$$dg_{2}(s)=b(g_{2}(s),u_{2}(s))ds+\sigma(g_{2}(s),u_{2}(s))dB^{G}(s),$$&\mbox{}\\ $$g_{2}(t+h)=g_{1}(t+h).$$&\mbox{}\end{cases}$$ (3.5)

Now define the control

$$u_{3}(s):=\begin{cases}$$u_{1}(s),$$&\mbox{$t\leq s\leq t+h$},\\ $$u_{2}(s),$$&\mbox{$t+h\leq s\leq T$},\end{cases}$$ (3.6)

and let

$$\begin{cases}$$dg_{3}(s)=b(g_{3}(s),u_{3}(s))ds+\sigma(g_{3}(s),u_{3}(s))dB^{G}(s),$$&\mbox{}\\ $$g_{3}(t)=g.$$&\mbox{}\end{cases}$$ (3.7)

By the pathwise uniqueness of strong solutions to SDE (3.1), we have

$$g_{3}(s)=\begin{cases}$$g_{1}(s),$$&\mbox{$t\leq s\leq t+h$},\\ $$g_{2}(s),$$&\mbox{$t+h\leq s\leq T$}.\end{cases}$$ (3.8)

By definition of the value function

$$W(t,g) \leq \mathbb{E}_{t,g}\left[\int_{t}^{T}L(g_{3}(s),u_{3}(s))ds+\Psi(g_{3}(T))\right]$$

$$\leq \mathbb{E}_{t,g}\left[\int_{t}^{t+h}L(g_{1}(s),u_{1}(s))ds+\int_{t+h}^{T}L(g_{2}(s),u_{2}(s))+\Psi(g_{2}(T))\right]$$

$$\leq \mathbb{E}_{t,g}\int_{t}^{t+h}L(g_{1}(s),u_{1}(s))ds+W(t+h,g_{1}(t+h))+\varepsilon.$$

Due to arbitrary of $\tilde{u}_{1}(\cdot)$, we obtain

$$W(t,g)\leq\inf\limits_{\mathcal{U}}\mathbb{E}_{t,g}\int_{t}^{t+h}L(g(s),u(s))ds+W(t+h,g(t+h))+\varepsilon,$$

where $g(\cdot)=g^{\tilde{u}(\cdot)}(\cdot)$. Since $\varepsilon>0$ is arbitrary, we prove the part “$\leq$”.

Next, we prove the part “$\geq$”. Fix $\varepsilon>0$, and select $u_{4}\in\mathcal{U}$ be such that

$$W(t,g)+\varepsilon\geq\mathbb{E}_{t,g}\left[\int_{t}^{T}L(g_{4}(s),u_{4}(s))ds+\Psi(g_{4}(T))\right],$$

where

$$\begin{cases}$$dg_{4}(s)=b(g_{4}(s),u_{4}(s))ds+\sigma(g_{4}(s),u_{4}(s))dB^{G}(s),$$&\mbox{}\\ $$g_{4}(t)=g.$$&\mbox{}\end{cases}$$

By definition of value function

$$W(t+h,g_{4}(t+h))\leq\mathbb{E}_{t+h,g_{4}(t+h)}\left[\int_{t+h}^{T}L(g_{4}(s),u_{4}(s))ds+\Psi(g_{4}(T))\right].$$

Therefore

$$W(t,g)+\varepsilon \geq \mathbb{E}_{t,g}\left[\int_{t}^{t+h}L(g_{4}(s),u_{4}(s))ds+\int_{t+h}^{T}L(g_{4}(s),u_{4}(s))+\Psi(g_{4}(T))\right]$$

$$\geq \mathbb{E}_{t,g}\left[\int_{t}^{t+h}L(g_{4}(s),u_{4}(s))ds\right]+\mathbb{E}_{t,g}W(t+h,g_{4}(t+h)).$$

Hence

$$W(t,g)+\varepsilon\geq\inf\limits_{\mathcal{U}}\mathbb{E}_{t,g}\left[\int_{t}^{t+h}L(g(s),u(s))ds+W(t+h,g(t+h))\right],$$

where $g(\cdot)=g^{u}(\cdot)$ is the strong solution of (3.4). The proof is completed.   

Now we derive the dynamic programming equation from (3.3) following standard argument as used in [21, 38].

Theorem 3.3

Let $\Psi\in C(G,\mathbb{R})$ and $W(T,g)=\Psi(g)$, $g\in G$. Assume that $W\in C^{1,2}([0,T]\times G,\mathbb{R})$ and satisfies Bellman’s principle of dynamic programming, that is, for any $0<t<\eta<T$ and $g\in G$,

$$W(t,g)=\inf\limits_{u\in\mathcal{U}}\mathbb{E}\left[\left.\int_{t}^{\eta}L(s,g(s),u(s))ds+W(\eta,g(\eta))\right|g(t)=g\right].$$

Then $W$ is a solution of the Bellman dynamic programming equation

$$0 = \inf\limits_{v\in\mathcal{U}}\left\{\partial_{t}W(t,g)+\langle\langle\nabla W(t,g),b(t,g,v)\rangle\rangle\right.$$

$$\left.+\frac{1}{2}{\rm Tr}\left[(\sigma(t,g,v)Q^{\frac{1}{2}})^{*}(\sigma(t,g,v)Q^{\frac{1}{2}})\nabla^{2}W(t,g)\right]+L(t,g,v)\right.\}.$$

Proof. The proof consists of two steps.

First step, we prove that $W$ is a sub solution. Choose $v\in\mathcal{U}$, we consider the constant control $u(s)\equiv v\in\mathcal{U},~{}t\leq s\leq t+h$. Let $\{g(s)\}$ be a strong solution of $(\ref{SDEG-1})$ driven by $u$. Then for all $h\in(0,T-t)$,

$$W(t,g)\leq\mathbb{E}\left[\int_{t}^{t+h}L(s,g(s),v)ds+W(t+h,g(t+h))|g(t)=g\right].$$

Subtracting $W(t,g)$ from both sides, dividing by $h$, and according to Itô-Dynkin formula , we obtain

$$0 \leq \frac{1}{h}\mathbb{E}\left[W(t+h,g(t+h))-W(t,g)|g(t)=g\right]$$

$$+\frac{1}{h}\mathbb{E}\left[\left.\int_{t}^{t+h}L(s,g(s),v)ds\right|g(t)=g\right]$$

$$\leq \frac{1}{h}\mathbb{E}\left[\int_{t}^{t+h}[\partial_{s}W(s,g(s))+\langle\nabla W(s,g(s)),b(s,g(s),v)\rangle\right.$$

$$+\frac{1}{2}{\rm Tr}[(\sigma(s,g(s),v)Q^{\frac{1}{2}})(\sigma(s,g(s),v)Q^{\frac{1}{2}})^{*}\nabla^{2}W(s,g(s))]$$

$$\left.\left.+L(s,g(s),v)ds]\right|g(t)=g\right].$$

Let $h\rightarrow 0$, for $v\in\mathcal{U}$, $W\in C^{1,2}([0,T]\times G;\mathbb{R})$, we have

$$0 \leq \partial_{t}W(t,g)+\langle\nabla W(t,g),b(t,g,v)\rangle$$

$$+\frac{1}{2}{\rm Tr}[(\sigma(t,g,v)Q^{\frac{1}{2}})(\sigma(t,g,v)Q^{\frac{1}{2}})^{*}\nabla^{2}W(t,g)]+L(t,g,v).$$

Then $W(t,g)$ is a sub solution of $(\ref{HJB-3})$ with terminal condition $W(T,g)=\Psi(g)$.

Second step, we prove that $W$ is a super solution.

For $\varepsilon\in(0,T-t)$, by definition of infimum, we can choose the minimizing sequences $u_{\varepsilon}\in\mathcal{U}$ such that

$$W(t,g)+\varepsilon^{2}\geq\mathbb{E}\left[\left.\int_{t}^{t+h}L(s,g^{\varepsilon}(s),u_{\varepsilon})ds+W(t+h,g^{\varepsilon}(t+h))\right|g^{\varepsilon}(t)=g\right].$$

We subtract $W(t,g)$ from bothsides, then divide by $\varepsilon$,

$$\varepsilon \geq \frac{1}{\varepsilon}\mathbb{E}\left[\left.W(t+\varepsilon,g^{\varepsilon}(t+\varepsilon))-W(t,g)\right|g^{\varepsilon}(t)=g\right]$$

$$+\frac{1}{\varepsilon}\mathbb{E}\left[\left.\int_{t}^{t+\varepsilon}L(s,g^{\varepsilon}(s),u^{\varepsilon})ds\right|g^{\varepsilon}(t)=g\right]$$

$$= \frac{1}{\varepsilon}\mathbb{E}\left[\int_{t}^{t+\varepsilon}[\partial_{s}W(s,g^{\varepsilon}(s))+\langle\nabla W(s,g^{\varepsilon}(s)),b(s,g^{\varepsilon}(s),u^{\varepsilon})\rangle\right.$$

$$+\frac{1}{2}{\rm Tr}[(\sigma(s,g^{\varepsilon}(s),u^{\varepsilon})Q^{\frac{1}{2}})(\sigma(s,g^{\varepsilon}(s),u^{\varepsilon})Q^{\frac{1}{2}})^{*}\nabla^{2}W(s,g^{\varepsilon}(s))]$$

$$\left.\left.+L(s,g^{\varepsilon}(s),g^{\varepsilon})ds]\right|g^{\varepsilon}(t)=g\right].$$

Let $\varepsilon\rightarrow 0$. Since $W\in C^{1,2}([0,T]\times G,\mathbb{R})$, noticing that $g^{\varepsilon}$ is the strong solution driven by the minimizing sequences, we obtain

$$0 \geq \partial_{t}W(t,g)+\inf\limits_{v\in\mathcal{U}}[\langle\langle\nabla W(t,g),b(t,g,v)\rangle\rangle$$

$$+\frac{1}{2}{\rm Tr}[(\sigma(t,g,v)Q^{\frac{1}{2}})(\sigma(t,g,v)Q^{\frac{1}{2}})^{*}\nabla^{2}W(t,g)]+L(t,g,v)].$$

Hence $W(t,g)$ is a super solution of $(\ref{HJB-3})$ with terminal condition $W(T,g)=\Psi(g)$. This finishes the proof.   

Remark 3.4

Notice that the dynamic programming equation is equivalent to

$$0=\min\limits_{v\in\mathcal{U}}\left[A^{v}W(t,g)+L(t,g,v)\right],$$ (3.10)

with the terminal condition

$$W(T,g)=\Psi(g),$$ (3.11)

where

$$A^{v}=\partial_{t}+b(t,g,v)\cdot\nabla+\frac{1}{2}{\rm Tr}[(\sigma(t,g,v)Q^{\frac{1}{2}})(\sigma(t,g,v)Q^{\frac{1}{2}})^{*}\nabla^{2}].$$ (3.12)

To simplify the notation, for fixed $Q$, we introduce the $Q$-Laplacian on $G$ as follows

$$\Delta_{Q}^{G}=\frac{1}{2}{\rm Tr}[(\sigma(t,g,v)Q^{\frac{1}{2}})^{*}(\sigma(t,g,v)Q^{\frac{1}{2}})\nabla^{2}].$$ (3.13)

Let $V\in C^{1}(G,\mathbb{R})$. Consider the Lagrangian action functional

$$L(s,g(s),u(s))=\frac{1}{2}\|u(s)\|_{T_{g(s)}G}^{2}-V(g(s)),$$

where $\|\cdot\|_{T_{g(s)}G}$ is the right-invariant $L^{2}$-Riemannian metric on $G={\rm SDiff}(M)$. By the dynamic programming equation $(\ref{HJB-4})$ derived from Theorem 3.3, we have

$$0 = \inf_{v\in\mathcal{U}}\left\{\partial_{t}W+\nu\Delta_{Q}^{G}W+v\cdot\nabla^{G}W+\frac{1}{2}\|v\|_{\mathcal{G}}^{2}-V(g)\right\}$$

$$= \partial_{t}W+\nu\Delta_{Q}^{G}W-\frac{1}{2}\|\nabla^{G}W\|_{\mathcal{G}}^{2}-V(g),$$

with terminal condition (3.11). Moreover, the optimal control $u$ is given by

$$u(t,g)=-\nabla^{G}W(t,g).$$ (3.15)

where $\nabla^{G}$ is the gradient on $G$, and $\Delta_{Q}^{G}$ is the $Q$-Laplacian on $G$.

Finally, let us mention that, without any difficulty we can extend the main result in this section to any $M$-type $2$ Banach manifold.

4 New derivation of the incompressible Navier-Stokes equations

In this section, we first derive the Hamilton-Jacobi-Bellman and the Burgers equations on the infinite dimensional group of volume preserving diffeomorphisms. By pushing forward to the underlying manifold, and using the Hodge decomposition theorem, we then derive the incompressible Navier-Stokes equations on compact Riemannian manifolds.

4.1 The idea for the derivation of incompressible Navier-Stokes equations

In Section $2$, we derive the Burgers equation $(\ref{M-Burgers})$ on compact Riemannian manifolds from the Bellman dynamic programming principle, which, at least in the case of bounded domain in Euclidean space, has the similar form as the incompressible Navier-Stokes equations except the divergence-free constraint condition ${\rm div}\hskip 0.85358ptu=0$. In a series works of Cruzeiro and her collaborators [10, 13], they developed a stochastic variational principle for the action

$$\mathcal{A}(x^{u},u) = \mathbb{E}\left[\int_{0}^{T}\frac{1}{2}\|D_{t}x^{u}(t)\|_{L^{2}(M)}^{2}dt\right]$$ (4.1)

$$= \mathbb{E}\left[\int_{0}^{T}\int_{M}\frac{1}{2}|D_{t}x^{u}(t)(x)|^{2}d\mu dt\right]$$

among the diffusion processes $\{x^{u}(t),t\in[0,T]\}$ defined by SDE $(\ref{SDEM})$ with control $u\in\mathcal{U}_{ad}$. More precisely, Cipriano and Cruzeiro [10] proved that the optimal control $u\in C^{1,2}([0,T]\times M,\Gamma(TM))$ for the Euler-Lagrangian equation of the Lagrangian action $\mathcal{A}$ on the two-dimensional torus $\mathbb{T}^{2}$ satisfies the incompressible Navier-Stokes equation

$$\partial_{t}u+u\cdot\nabla u{\color[rgb]{0,0,1}-}\nu\Delta u = -\nabla p,$$

$${\rm div}\hskip 0.85358ptu = 0.$$

See further works in [2, 3, 19, 29] for the extension of this idea to compact Riemannian manifolds.

Let $\mathcal{V}:G\rightarrow\mathbb{R}$ be a $C^{1}$-continuous function. We introduce

$$\mathcal{L}(g^{u},u)=\mathbb{E}\left[\int_{0}^{T}\left(\int_{M}\frac{1}{2}|D_{s}g^{u}(s)(x)|^{2}d\mu-\mathcal{V}(g^{u}(s))\right)ds\right],$$ (4.2)

which is the Lagrangian action function on $G={\rm SDiff}(M)$, where $g^{u}(s)$ is the horizontal lift of the solution $x^{u}(s)$ of the Controlled SDE $(\ref{SDEM})$ on $G={\rm SDiff}(M)$.

Let $\Psi\in C(G,\mathbb{R})$. We introduce the value function $W$ on $G={\rm SDiff}(M)$

$$W(t,g)=\inf\limits_{\mathcal{U}}\mathbb{E}_{t,g}\left[\int_{t}^{T}\left(\int_{M}\frac{1}{2}|D_{s}g^{u}(s)(x)|^{2}d\mu-\mathcal{V}(g^{u}(s))\right)ds+\Psi(g^{u}(T))\right].$$ (4.3)

Here and throughout this paper, $\mathbb{E}_{t,g}[\ldots]=\mathbb{E}\left[\left.\ldots\right|g^{u}(t)=g\right]$.

By the Bellman dynamic programming principle on $G={\rm SDiff}(M)$ which has been proved in Section $3$, we have

$$0 = \inf_{v\in T_{g}G}\left\{\partial_{t}W+\nu\Delta^{G}W+\langle v,\nabla^{G}W\rangle_{T_{g}G}+\frac{1}{2}\|v\|_{T_{g}G}^{2}-\mathcal{V}(g)\right\}$$

$$= \partial_{t}W+\nu\Delta^{G}W+\inf_{v\in T_{g}G}\left\{\langle v,\nabla^{G}W\rangle_{T_{g}G}+\frac{1}{2}\|v\|_{T_{g}G}^{2}\right\}-\mathcal{V}(g).$$

where $\nabla^{G}$ and $\Delta^{G}$ denote the Riemannian gradient operator and the weighted Laplace-Beltrami (i.e., the Ornstein-Uhlenbeck) type operator on the infinite dimensional Riemannian manifold on $G={\rm SDiff}(M)$.

Then, by a similar argument as used in the derivation of the Burgers equation $(\ref{M-Burgers})$ on $M$, we can prove that the value function $W$ satisfies the infinite dimensional Hamilton-Jacobi-Bellman equation on $G={\rm SDiff}(M)$

$$\partial_{t}W+\nu\Delta^{G}W-\frac{1}{2}\|\nabla^{G}W\|^{2}_{T_{g}G}-\mathcal{V}(g)=0,$$ (4.5)

with the terminal condition $W(T,g)=\Psi(g)$. Moreover, the optimal Markov control $U$ is given by

$$U=-\nabla^{G}W.$$

By the same argument as explained in the derivation of the Burgers equation $(\ref{M-Burgers})$, we will prove that $U$ satisfies the Burgers type equation on $G$.

Furthermore, by the Hodge decomposition theorem, and by pushing forward $(\ref{HJBG})$ to $M$ via the evaluation map $e_{x}:G\rightarrow M,g\mapsto g(x)$, we can derive that $u=(de_{x})U$ satisfies the incompressible Navier-Stokes equation on the underlying Riemannian manifold $(M,g)$.

This gives a new derivation of the incompressible Navier-Stokes equation via the Bellman dynamic programming principle on the infinite dimensional group $G={\rm SDiff}(M)$. In particular, when the viscosity coefficient $\nu$ vanishes, we give a new derivation of the incompressible Euler equation via the deterministic dynamic programming principle on the infinite dimensional group $G={\rm SDiff}(M)$.

4.2 The Hamilton-Jacobi-Bellman equation on $G={\rm SDiff}(M)$

In this subsection, we introduce a controlled SDE on the group of volume preserving diffeomorphisms $G={\rm SDiff}(M)$ and derive the Hamilton-Jacobi-Bellman equation on $G={\rm SDiff}(M)$.

Let $\{g^{u}(t),t\in[0,T]\}$ be a diffusion process with values in $G={\rm SDiff}(M)$ and satisfies the following Stratonovich stochastic differential equation (briefly, SDE) with right action on $G={\rm SDiff}(M)$

$$\begin{cases}$$dg^{u}(t)=\left(\sqrt{2\nu}\circ dW_{t}\right)g^{u}(t)+u(t,g^{u}(t))dt,$$&\mbox{$$}\\ $$g^{u}(0)=e,$$&\mbox{$$}\end{cases}$$ (4.6)

where $W_{t}$ is a $\mathcal{G}$-valued $Q$-Brownian motion (see Definition 3.1), $u$ is a horizontal lift of a smooth divergence-free vector field on $M$ to $\mathcal{G}$, and $e=id$ is the identity map of $G={\rm{SDiff}}(M)$. The short time existence and uniqueness of solution to SDE $(\ref{SDEG})$ follows from standard Picard iteration as in Itô’s SDE theory on Euclidean spaces.

Let $c_{i}\in\mathbb{R}$, $\{W_{i},i=1,2,...\}$ be i.i.d. Brownian motions, $\{\widetilde{A}_{i},i=1,2,...\}$ and $U$ be smooth vector fields on $G={\rm SDiff(M)}$. Let $g(s)$ be the solution to the following SDE on $G$

$$\begin{cases}$$dg(s)=\sqrt{2\nu}\sum\limits_{i=1}^{\infty}c_{i}\tilde{A}_{i}(g(s))\circ dW_{i}(s)+U(s,g(s))ds,$$&\mbox{$s\in(t,T]$}\\ $$g(t)=g.$$&\mbox{}\end{cases}$$ (4.7)

Since $\{\widetilde{A}_{i},i=1,2,...\}$ and $U$ are smooth vector fields on $G$, SDE $(\ref{ch2-SDEG})$ has a unique solution $\{g(s),s\in[t,T]\}$.

Let $e_{x}$ be the evaluation map, $e_{x}:G\rightarrow M,~{}~{}e_{x}(g)=g(x)$,   $de_{x}$ be the differential of $e_{x}$ at the point $g$

$$de_{x}:T_{g}G\rightarrow T_{g(x)}M.$$

By the chain rule of differential, we have

$$d(e_{x}\circ R_{g})_{e}=(de_{x})_{g}(dR_{g})_{e}.$$

Lemma 4.1

Suppose that there exists a family of vector fields $\{A_{i}\}_{i\in\mathcal{I}}$ on $M$ such that for all $g\in G$, it holds

$$A_{i}(g(x))=d(e_{x}\circ R_{g})_{e}\widetilde{A_{i}}(e),$$

where $\widetilde{A_{i}}(e)\in\mathcal{G}$ is the horizontal lift of $A_{i}(x)$ by $e_{x}$, and is right-invariant, i.e.,

$$\widetilde{A_{i}}(g)=(dR_{g})_{e}\widetilde{A_{i}}(e).$$

Then for any $f\in C^{\infty}(M,\mathbb{R})$, and $F_{x}=f\circ e_{x}$, it holds

$$(\widetilde{\mathcal{L}}F_{x})(g)=(\mathcal{L}f)(g(x)),$$

where $\widetilde{\mathcal{L}}$ is the infinitesimal generator of SDE on $G$

$$dg_{t}=\sum c_{i}\widetilde{A}_{i}(g_{t})\circ dW_{t}^{i}+\widetilde{A}_{0}(g_{t})dt,~{}~{}g_{0}=g,$$

and $\mathcal{L}$ is the infinitesimal generator of SDE on $M$

$$dg_{t}(x)=\sum c_{i}A_{i}(g_{t}(x))\circ dW_{t}^{i}+A_{0}(g_{t}(x))dt,~{}~{}g_{0}(x)=g(x)$$

Proof. By the fact $A_{i}(g(x))=d(e_{x}\circ R_{g})_{e}\widetilde{A_{i}}(e)=(de_{x})_{g}(dR_{g})_{e}\widetilde{A_{i}}(e)=(de_{x})_{g}\widetilde{A_{i}}(g)$, we have $\widetilde{A_{i}}F_{x}(g)=(A_{i}f)(g(x))=(A_{i}f)(e_{x}(g))=(A_{i}f)\circ e_{x}(g)$. Hence

$$\widetilde{A_{i}}F_{x} = (A_{i}f)\circ e_{x},$$

$$\widetilde{A_{i}}(\widetilde{A_{i}}F_{x}) = A_{i}(A_{i}f)\circ e_{x}.$$

This implies

$$(\widetilde{\mathcal{L}}F_{x})(g) = \left.\frac{\partial}{\partial t}\mathbb{E}_{g}[F_{x}(g_{t})]\right|_{t=0}$$

$$= \sum c_{i}^{2}\widetilde{A_{i}}^{2}F_{x}(g)+\widetilde{A_{0}}F_{x}(g)\cdot\nabla$$

$$= \sum c_{i}^{2}A_{i}^{2}f(g(x))+A_{0}f(g(x))\cdot\nabla$$

$$= \left.\frac{\partial}{\partial t}\mathbb{E}_{g(x)}[f(g_{t}(x))]\right|_{t=0}$$

$$= (\mathcal{L}f)(g(x)).$$

 

Fix $x\in M$. Let $e_{x}:G\rightarrow M$ be the evaluation map, and $g(s)(x)=e_{x}(g(s))$. By the Itô formula, we can derive the SDE for $g(s)(x)$ on $M$ from SDE $(\ref{ch2-SDEG})$ for $g(s)$ on $G$ as follows

$$\begin{cases}$$dg(s)(x)=\sqrt{2\nu}\sum\limits_{i=1}^{\infty}c_{i}A_{i}(g(s)(x))\circ dW_{i}(s)+U(s,g(s)(x))ds,$$&\mbox{$s\in(t,T]$}\\ $$g(t)(x)=g(x).$$&\mbox{}\end{cases}$$ (4.8)

Here the vector fields $\tilde{A}_{i}$ and $U$ on $G={\rm SDiff(M)}$ and $A_{i}$ and $u$ on $M$ satisfy the relationship

$$(de_{x})(g)U(g)=u(g(x)),$$

$$(de_{x})(g)\tilde{A}_{i}(g)=A_{i}(g(x)),i=1,2,\cdots,$$

which is the case when $\widetilde{A}_{i}$ and $U$ are right-invariant on $G$ in the sense that

$$U(g)=dR_{g}U(e),$$

$$\tilde{A}_{i}(g)=dR_{g}\tilde{A}_{i}(e),\ \ i=1,2,\ldots$$

$$(de_{x})U(e)=u(x),$$

$$(de_{x})\widetilde{A}_{i}(e)=A_{i}(x),\ \ i=1,2,\ldots$$

Since $A_{i}$ and $u$ are $C^{\infty}$ vector field on $M$, then SDE $(\ref{ch2-SDED})$ has a unique solution $g(s)(x)$, and $x\mapsto g(s)(x)$ is a diffeomorphism on $M$ using the theory of stochastic flows. See [22] and [25]. The infinitesimal generator of $g(s)(x)$ of SDE $(\ref{ch2-SDED})$ is given by

$$L=\nu\sum\limits_{i=1}^{\infty}c_{i}^{2}A_{i}^{2}+u.$$

We need a family of vector fields $\{A_{i},i\in\mathbb{N}\}$ and renormalization coefficients $\{c_{i},.i\in\mathbb{N}\}$ such that

$$\sum\limits_{i=1}^{\infty}c_{i}^{2}A_{i}^{2} = c\Delta^{M},$$ (4.9)

$${\rm div}A_{i} = 0,\ \ \forall i\in\mathbb{N}$$ (4.10)

where $c>0$ is a constant, $\Delta^{M}$ is the Laplace-Beltrami operator on the Riemannian manifold $(M,g)$. When $M=\mathbb{T}^{d}$, this was proved by Cipriano-Cruzeiro [10]. When $M=\mathbb{S}^{d}$, this was proved by Le Jan-Raimond [27], see also Fang-Zhang [20], Fang-Luo [19] and Luo [29]. In [15], Elworthy, Le Jan and X.-M. Li proved the existence of such vector fields on compact Riemannian manifolds equipped with a Le Jan-Watanabe connection, in particular, for gradient Brownian system on compact Riemannian manifolds equipped with a torsion-skew symmetric metric connection. More precisely, assuming that there exist a family of vector fields $\{A_{i},i\in\mathbb{N}\}$ such that

$$\|v\|^{2}=\sum\limits_{i=1}^{\infty}\langle A_{i}(x),v\rangle^{2},\ \ \forall x\in M,v\in T_{x}M,$$ (4.11)

$$\sum\limits_{i=1}^{\infty}\nabla_{A_{i}}A_{i}=0,$$ (4.12)

$$\sum\limits_{i=1}^{\infty}A_{i}\wedge\nabla_{X}A_{i}=0,\ \ \forall X\in\Gamma(TM),$$ (4.13)

then for any differential form $\omega$ on $M$, it holds (see Remark 2.3.1 in [15])

$$\square^{M}\omega=-\sum\limits_{i=1}^{\infty}L_{A_{i}}^{2}\omega.$$ (4.14)

Moreover, we have the following

Lemma 4.2

([27, 10, 20, 19]) Assume that there exist a family of vector fields $\{A_{i},i\in\mathbb{N}\}$ and $\{c_{i},i\in\mathbb{N}\}$ satisfying $(\ref{cA1})$ and $(\ref{cA2})$. Fix $x\in M$. Let $e_{x}:G\rightarrow M$ be the evaluation map $e_{x}(g)=g(x)$. Then, for any $f\in C^{\infty}(M)$, it holds

$$L_{G}(f\circ e_{x})=(L_{M}f)\circ e_{x}.$$ (4.15)

Equivalently, letting $F_{x}:G\rightarrow\mathbb{R}$ be defined by

$$F_{x}(g)=(f\circ e_{x})(g)=f(g(x)),\ \ \forall g\in G.$$

Then

$$(L_{G}F_{x})(g)=(L_{M}f)(g(x)).$$ (4.16)

The main result of this subsection is the following Hamilton-Jacobi-Bellman equation on $G=\rm{SDiff}(M)$.

Theorem 4.3

Let $g(s)$ be the solution to the controlled SDE $(\ref{SDEG})$. Let

$$W(t,g):=\inf\limits_{U\in\mathcal{U}}\mathbb{E}_{t,g}\left[\int_{t}^{T}L(s,g(s),U(s))ds+\Psi(g(T))\right]$$ (4.17)

be the value function of the related stochastic control problem, where the Lagrangian is given by

$$L(s,g(s),U(s))=\frac{1}{2}\|U\|_{T_{g(s)}G}^{2}-V(g(s)).$$

Then $W$ satisfies the Hamilton-Jacobi-Bellman equation on $G=\operatorname{SDiff}(M)$

$$\partial_{t}W-\frac{1}{2}\|\nabla^{G}W\|^{2}_{T_{g}G}+\nu\Delta^{G}W=V(g),$$ (4.18)

with the terminal condition $W(T,g)=\Psi(g)$.

Proof. By the extended Bellman dynamic programming principle on $G$ (see Eq. $(\ref{HJB-4})$ in Section $3$), we have

$$\inf_{U\in\mathcal{G}}\{\partial_{t}W+U\cdot\nabla^{G}W+\nu\Delta^{G}W+L(t,g,U)\}=0.$$

By the same argument as used in Section 2 for the proof of the HJB equation on $M$, see also the description in Subsection 2.1, we conclude that the optimal Markov control $U(t,g)$ is given by

$$U(t,g)=-\nabla^{G}W(t,g),$$

and $W$ satisfies the Hamilton-Jacobi-Bellman equation $(\ref{HJB-W})$ with the terminal condition $W(T,g)=\Psi(g)$.   

4.3 New derivation of the incompressible Navier-Stokes equation

In this subsection, we use the Hamilton-Jacobi-Bellman equation on $G=\operatorname{SDiff}(M)$ to give a new derivation of the incompressible Navier-Stokes equation on a compact Riemannian manifold $M$.

Let $d^{G}$ be the exterior differential on $G$. Following [15], we introduce

$$\delta^{G}=\sum\limits_{i=1}^{\infty}c_{i}^{2}{\rm int}_{\widetilde{A}_{i}}\nabla_{\widetilde{A}_{i}},$$

and define the Hodge like Laplacian acting on $1$-form on $G$ as follows

$$\square^{G}=d^{G}\delta^{G}+\delta^{G}d^{G}.$$

Since $(d^{G})^{2}=0$, it holds

$$d^{G}\square^{G} = d^{G}d^{G}\delta^{G}+d^{G}\delta^{G}d^{G}$$

$$= d^{G}\delta^{G}d^{G}$$

$$= d^{G}\delta^{G}d^{G}+\delta^{G}d^{G}d^{G}$$

$$= \square^{G}d^{G}.$$

Now we state the main result of this paper as follows.

Theorem 4.4

Let $W\in C^{1,3}([0,T]\times G)$ be a solution to the Hamilton-Jacobi-Bellman equation $(\ref{HJB-W})$ on $G={\rm SDiff}(M)$. Let

$$U(t,g)=-\nabla^{G}W(t,g)$$

be the optimal control. Then

$$\partial_{t}U+\frac{1}{2}\nabla^{G}\|U\|^{2}-\nu\Delta^{G}U=-\nabla^{G}V.$$

Let

$$u(t,x)=(de_{x})U(t,e),~{}v(x)=(de_{x})V(e).$$

Then there exists a function $p\in C^{1}([0,T]\times M,\mathbb{R})$ such that $(u,p)$ satisfies the backward incompressible Navier-Stokes equation on $M$

$$\partial_{t}u+\nabla^{TM}_{u}u-\nu\square^{M}u = -\nabla^{M}p,$$ (4.19)

$$\rm{div}\hskip 0.85358ptu = 0,$$ (4.20)

and $(-u(T-t,\cdot),p(T-t,\cdot))$ satisfies the incompressible Navier-Stokes equation on $M$

$$\partial_{t}u+\nabla^{TM}_{u}u+\nu\square^{M}u = -\nabla^{M}p,$$ (4.21)

$$\rm{div}\hskip 0.85358ptu = 0.$$ (4.22)

Note that the $\nabla^{M}v$ has been absorbed into $\nabla^{M}p$.

Proof. Taking differentiation on the Hamilton-Jacobi-Bellman equation $(\ref{HJB-W})$, we have

$$d^{G}(\partial_{t}W)-d^{G}\left(\frac{1}{2}\|\nabla^{G}W\|^{2}_{T_{g}G}\right)+\nu d^{G}(\Delta^{G}W)=d^{G}V(t,g).$$

Using the commutative formulae

$$d^{G}\partial_{t}W=\partial_{t}d^{G}W,\ d^{G}\Delta^{G}W=-\square^{G}d^{G}W,$$

we have

$$\partial_{t}\nabla^{G}W-\frac{1}{2}\nabla^{G}\|\nabla^{G}W\|^{2}_{T_{g}G}-\nu\square^{G}\nabla^{G}W=\nabla^{G}V(t,g).$$ (4.23)

Thus $U(t,g)=-\nabla^{G}W(t,g)$ satisfies the Burgers equation on $G$

$$\partial_{t}U+\frac{1}{2}\nabla^{G}\|U\|^{2}_{T_{g}G}-\nu\square^{G}U=-\nabla^{G}V(t,g).$$

Pushing forward the Burgers equation on $G$ to $M$ by the evaluation map

$$e_{g^{-1}(x)}:G=\operatorname{SDiff}(M)\rightarrow M,\quad g\mapsto x.$$

By definition of the right invariant vector fields on $G$, there is a one to one correspondence between vector fields $u,~{}v$ on $M$ and right invariant vector fields $U,~{}V$ on $G=\operatorname{SDiff}(M)$ such that

$$u(t,x)=\left(de_{g^{-1}(x)}\right)U(t,g),~{}v(x)=\left(de_{g^{-1}(x)}\right)V(g).$$

By the commutative formula $d\Delta^{G}W=-\square^{G}dW$ and the duality between the Riemmanian gradient and exterior differential operator on $\mathcal{G}=\{X\in C^{\infty}(M,TM):\rm{div}\hskip 0.85358ptX=0\}$, for any $X\in\mathcal{G}=T_{e}G$, we have

$$\langle\langle\nabla^{G}(\Delta^{G}W),X\rangle\rangle = \langle\langle\Box^{G}U,X\rangle\rangle=\int_{M}\langle\Box^{M}u(t,x),X(x)\rangle d\mu(x).$$ (4.24)

On the other hand, it holds

$$\left\langle\left\langle\nabla^{G}(\partial_{t}W),X\right\rangle\right\rangle = -\int_{M}\langle\partial_{t}u(t,x),X(x)\rangle d\mu(x),$$ (4.25)

$$\left\langle\left\langle\nabla^{G}\left(\frac{1}{2}\|\nabla^{G}W\|^{2}_{T_{g}G}\right),X\right\rangle\right\rangle = \int_{M}\left\langle\nabla^{M}\left(\frac{1}{2}|u(t,x)|^{2}\right),X(x)\right\rangle d\mu(x),$$ (4.26)

$$\left\langle\left\langle\nabla^{G}V,X\right\rangle\right\rangle = \int_{M}\left\langle\nabla^{M}v(t,x),X(x)\right\rangle d\mu(x).$$ (4.27)

Combining $(\ref{DSW2})$, $(\ref{DSW3})$, $(\ref{DSW4})$ and $(\ref{DSW4v})$ with $(\ref{HJBW})$, we have

$$0 = \left\langle\left\langle\nabla^{G}(\partial_{t}W)-\nabla^{G}\left(\frac{1}{2}\|\nabla^{G}W\|^{2}_{T_{g}G}\right)+\nu\nabla^{G}(\Delta^{G}W)-\nabla^{G}V(t,g),X\right\rangle\right\rangle$$

$$= \int_{M}\left\langle-\partial_{t}u(t,x)-\frac{1}{2}\nabla^{M}|u(t,x)|^{2}+\nu\Box^{M}u(t,x)-\nabla^{M}v(t,x),X(x)\right\rangle d\mu(x).$$

By the Hodge decomposition theorem, there exists a function $p\in C^{1}([0,T]\times M)$ such that $u=u(t,x)$ satisfies the backward incompressible Navier-Stokes equation on M

$$\begin{cases}$$\partial_{t}u+\nabla^{TM}_{u}u-\nu\Box^{M}u=-\nabla^{M}p,$$&\mbox{}\\ $$\rm{div}\hskip 0.85358ptu=0,$$&\mbox{}\end{cases}$$

where we use

$$\nabla^{M}|u|^{2}=2\nabla^{TM}_{u}u.$$

 

5 The incompressible Euler equation

In this section, we consider the particular case where the viscosity vanishes, i.e., $\nu=0$, and derive the incompressible Euler equation from the deterministic dynamic programming principle on the group $G={\rm SDiff}(M)$ over a compact Riemannian manifold.

Consider the deterministic control problem on $G={\rm SDiff}(M)$

$$W(t,g):=\inf\limits_{U\in\mathcal{U}}\left[\left.\int_{t}^{T}L(s,g(s),U(s))ds+\Psi(g_{T})\right|g(t)=g\right],$$

where $g(s)$ is the solution to the ODE on $G$

$$\begin{cases}$${d\over ds}g(s)=U(s,g(s)),$$&\mbox{$s\in(t,T],$}\\ $$g(t)=g.$$&\mbox{}\end{cases}$$ (5.1)

Consider the Lagrangian

$$L(s,g(s),U(s))=\frac{1}{2}\|U\|_{T_{g(s)}G}^{2}-V(g(s)).$$

By the extended deterministic dynamic programming principle on $G={\rm SDiff}(M)$, whose proof can be given similarly to the proof of Theorem 3.2, we have

$$\inf_{U\in T_{g}G}\{\partial_{t}W+U\cdot\nabla^{G}W+L(t,g,U)\}=0.$$

The optimal Markov control $U(t,g)$ is of the form

$$U(t,g)=-\nabla^{G}W(t,g),$$

and the value function $W$ satisfies the Hamilton-Jacobi equation

$$\partial_{t}W-\frac{1}{2}\|\nabla^{G}W\|^{2}_{T_{g}G}=V(t,g),$$ (5.2)

with the terminal value

$$W(T,g)=\Psi(g).$$

The following result gives a new derivation of the incompressible Euler equation from the Hamilton-Jacobi equation on $G={\rm SDiff}(M)$.

Theorem 5.1

Let $W\in C^{1,2}([0,T]\times G)$ be a solution to the Hamilton-Jacobi equation $(\ref{HJ-W})$. Let

$$U(t,g)=-\nabla^{G}W(t,g)$$

be the optimal control and

$$u(t,x)=(de_{x})_{e}U(t,e).$$

Then there exists a function $p\in C^{1}(M,\mathbb{R})$ such that $(u,p)$ satisfies the incompressible Euler equation on $M$

$$\partial_{t}u+\nabla^{TM}_{u}u = -\nabla^{M}p,$$ (5.3)

$${\rm div}\hskip 0.85358ptu = 0.$$ (5.4)

Proof. The proof is similar to the one of Theorem 4.4 by taking $\nu=0$.   

6 Conclusion and some problems for further work

In conclusion, the main result of this paper can be formulated as follows.

Theorem 6.1

Let $M$ be a compact Riemannian manifold. Let $W$ be a $C^{1,3}$-smooth solution to the Hamilton-Jacobi-Bellman equation $(\ref{HJB-W})$ on $G={\rm SDiff}(M)$. Then

$$u(t,x)=(de_{x})\nabla^{G}W(t,e)$$

is a $C^{1,2}$-smooth solution to the backward incompressible Navier-Stokes equation $(\ref{NSM-u1})$. Conversely, let $u$ be a smooth solution to the backward incompressible Navier-Stokes equation $(\ref{NSM-u1})$. Let $\{g_{t},t\in[0,T]\}$ be the unique solution to SDE $(\ref{SDEG})$ with $g(0)=e$ on $G={\rm SDiff}(M)$. Then the value function $W$ for the optimal stochastic control problem $(\ref{W-Gvaluefunction})$ is a solution to the Hamilton-Jacobi-Bellman equation $(\ref{HJB-W})$ on $G={\rm SDiff}(M)$.

Proof. This is indeed the restatement of our main results in Section 2 and Section 4.   

To end this paper, we raise some problems for further research work.

Remark 6.2

Our main result indicates an interesting relationship between the incompressible Navier-Stokes equation on a compact Riemannian manifold and the Hamilton-Jacobi-Bellman equation on the group of volume preserving diffeomorphisms. By [14, 35], if the initial value $u_{0}\in C^{\infty}(M)$, then there exists some $T>0$ such that the incompressible Navier-Stokes equation $(\ref{NSMF-u1})$ has a local smooth solution $u\in C^{\infty}([0,T]\times M)$. This yields that the vector fields in SDE $(\ref{SDEG})$ are smooth on $G$ and hence SDE $(\ref{SDEG})$ admits a unique strong solution on $[0,T]\times G$. It is interesting to ask the question whether we can prove the regularity of the solution to the Hamilton-Jacobi-Bellman equation $(\ref{HJB-W})$ on the group of volume preserving diffeomorphisms.

By the Cole-Hopf transformation

$$\Phi(t,g)=\exp{\left[-{W(t,g)\over{2\nu}}\right]}$$

and its inverse transform, it is well-known that, $W$ is a solution to the Hamilton-Jacobi-Bellman equation $(\ref{HJB-W})$ on $G={\rm SDiff}(M)$ if and only if $\Phi$ is a solution to the heat equation

$$\partial_{t}\Phi+\nu\Delta^{G}\Phi+\frac{1}{2\nu}V\Phi=0.$$ (6.1)

It remains a challenge problem whether we can prove the existence and uniqueness of local or global smooth solution of the Hamilton-Jacobi-Bellman equation $(\ref{HJB-W})$ and the heat equation $(\ref{HeatVG})$ on $G={\rm SDiff}(M)$. Instead of smooth solution, it is also interesting to study viscosity solution or solution in suitable Sobolev space in the sense of the Malliavin calculus [30] over $G={\rm SDiff}(M)$. This will be our future research project.

Acknowledgement. The authors would like to express their gratitudes to Professors A. B. Cruzeiro, F. Flandoli and T. Funaki for their interests and stimulating discussions in the beginning step of this work. We also thank Dr. S. Li for helpful discussions in the preparation of this work.

Guo** Liu, School of Mathematics and Statistics, Huazhong University of Science and Technology, Wuhan, No.1037, Luoyu Road, Wuhan, 430074, China

E-mail: [email protected]

Xiang-Dong Li, Academy of Mathematics and Systems Science, Chinese Academy of Sciences, No. 55, Zhongguancun East Road, Bei**g, 100190, China

E-mail: [email protected]

and

School of Mathematical Sciences, University of Chinese Academy of Sciences, Bei**g, 100049, China